AMD Considered GDDR5 For Kaveri, Might Release Eight-Core Variant

MojoKid writes "Of all the rumors that swirled around Kaveri before the APU debuted last week, one of the more interesting bits was that AMD might debut GDDR5 as a desktop option. GDDR5 isn't bonded in sticks for easy motherboard socketing, and motherboard OEMs were unlikely to be interested in paying to solder 4-8GB of RAM directly. Such a move would shift the RMA responsibilities for RAM failures back to the board manufacturer. It seemed unlikely that Sunnyvale would consider such an option but a deep dive into Kaveri's technical documentation shows that AMD did indeed consider a quad-channel GDDR5 interface. Future versions of the Kaveri APU could potentially also implement 2x 64-bit DDR3 channels alongside 2x 32-bit GDDR5 channels, with the latter serving as a framebuffer for graphics operations. The other document making the rounds is AMD's software optimization guide for Family 15h processors. This guide specifically shows an eight-core Kaveri-based variant attached to a multi-socket system. In fact, the guide goes so far as to say that these chips in particular contain five links for connection to I/O and other processors, whereas the older Family 15h chips (Bulldozer and Piledriver) only offer four Hypertransport links."

Comments

Latency vs bandwidth

[evilsofa]

DDR3 is low latency, low bandwidth. GDDR5 is high latency, high bandwidth. Low latency is critical for CPU performance while bandwidth doesn't matter as much. On video cards, GPUs need high bandwidth but the latency doesn't matter as much. This is why gaming PCs use DDR3 for system RAM and GDDR5 on their video cards. Video cards that cut costs by using DDR3 instead of GDDR5 take a massive hit in performance. The XBox One and PS4 use GDDR5 shared between the CPU and GPU, and as a result have the rough equivalent of a very low-end CPU paired with a mid-range GPU.

Re:Latency vs bandwidth

[symbolset]

We had this argument long ago, and decided we don't like Rambus not because they don't have some good tech, but because they are fucktards.

Re:Latency vs bandwidth

[dagamer34]

The Xbox One doesn't use GDDR5 but DDR3 memory with 32MB eSRAM. They made the decision early on to have 8GB of RAM (regardless of type) before they knew it would be feasible to have 8GB GDDR5 memory at an easily attainable cost.

Re:Latency vs bandwidth

[Anonymous Coward]

False, XBox one uses pure DDR3.

It is also one of the key reasons why many games on XBox one cannot do 1080p (that, and the lack of ROPs - PS4 having twice as many ROPs for rasterization)

XBox One tries to "fix" the RAM speed by using embedded sRAM on-chip as a cache for the DDR3 for graphics. Remains to be seen how well the limitations of DDR3 can be mitigated. Early games are definitely suffering from "developer cannot be assed to do a separate implementation for Xbox One".

Kaveri, while related to the chips inside the consoles, is decisively lower performing part. Kaveri includes 8 CUs. XBox one has 14 CUs on die, but two of those are disabled (to improve yields), so 12. PS4 has 20 CUs on-die, with again two CUs disabled to improve yields, so 18.

On the other hand, Kaveri has far better CPU cores (console chips feature fairly gimpy Jaguar cores, tho both consoles feature 8 of those cores vs 4 on Kaveri)

Any integrated graphics setup that uses DDR3 is bound to be unusable for real gaming. Kaveri has a good integrated graphics setup compared to the competition, but it is far behind what the new consoles feature - boosting it with GDDR5 without also at least doubling the CU count wouldn't do much. Either way, it really isn't usable for real gaming. It beats the current top offering from Intel, but that's bit like winning in Special Olympics when compared to real graphics cards (even ~$200-250 midrange ones)

Re:Latency vs bandwidth

[Anonymous Coward]

Somewhat false. Latency is approximately the same for DDR3 vs GDDR5, at least in terms of nanoseconds from request to response. GDDR5 is impractical for use in CPUs due to the need to solder to a board and high power consumption (enough to need cooling). That CPUs don't need so much bandwidth just makes it useless in addition.

Also, the Xbox One uses DDR3, and makes up for the lack of bandwidth with four channels and 32MB of ESRAM for graphics use.

Re:Latency vs bandwidth

[Sockatume]

Latency in cycles is higher for GDDR5, but the clock speed's a lot faster, isn't it? As the real-time latency is the product of the number of cycles and the length of a cycle, I think it's pretty much a wash.

Re:Latency vs bandwidth

[Bengie]

DDR3 and GDDR5 have nearly the same latency when measured in nanoseconds. When measured in clock cycles GDDR5 is higher latency, but it has more cycles. This has to do with the external interface, which is more serial in nature than the internal. The data path for the internal is quite wide, but the external datapath is not because traces on a motherboard are expensive. To crank up the bandwidth, you increase the frequency. If the internal frequency remains fixed and the external frequency goes up, the external "latency" seeming goes up.

Re:news for nerds

[ls671]

Great! Can you explain to me why "GDDR5 isn't bonded in sticks for easy motherboard socketing" ?

Of course, that would miss the point

[guacamole]

The whole point of AMD APUs is low cost gaming. That is, lower cost than buying a dedicated GPU plus a processor. Many already argue that you don't save much by buying an APU. A cheap Pentium G3220 with a AMD Radeon 7730 costs the same as the A10 Kaveri APU, and will give better frame rate. Even if the Kaveri APU prices come down, the savings will be small. If you have to buy the GDDR5 memory, there won't be any savings. It's understandable that AMD didn't take that route.

Re:Of course, that would miss the point

[girlintraining]

The whole point of AMD APUs is low cost gaming.

Sigh. You people with your myopic vision. If AMD consigned itself to your view of what it should do, it'll be dead in another 5-7 years. Let's take a look at what Intel offers: Higher performance. Lower energy consumption. Less heat. Smaller die size. In fact, you'd be hard pressed to find anything AMD has in its favor from an engineering standpoint. So what does AMD have that's keeping it in business? Cost. AMD offers a lower price point for economy systems.

But that's not where the profit is. That's not what's going to take AMD into the mid-21st century. If AMD sticks to that line of thinking, it'll go the way of Cyrix... and for exactly the same reason. AMD can't invest in a new fab plant because its cash reserves are too low, whereas Intel's pile of gold just keeps growing.

The only way to swing AMD back into the mid and high-end market is parallelization -- multiple chips, embedding everything into everything, and cutting costs everywhere it can... because it can't shrink the CPU and it can't do a damn thing about energy consumption. Their only avenue of escape is a radical re-think of the chipset, integration, and anything it can to boost profits to make that leap forward in minaturization to catch up with Intel. And it will not be easy. Many market analysts have suggested AMD throw in the towel.

That's the position AMD has to play. Of course, your myopic view only sees an engineering problem to solve. You have absolutely no comprehension of the business considerations behind the move... which is probably why you insist they're missing the boat. They are trying to claw their way into the mid-range market and undercut Intel. This is one of the select few ways they're going to do it. Unfortunately, the cost of commodity components to get this Hindenburg off the ground isn't going down -- AMD's position is shrinking and if they don't leverage everything now, they're going to find themselves in an untenable market position and will simply fold, leaving Intel as the only major player left in the market.

And if that happens... we're all fucked.

Re:Of course, that would miss the point

[edxwelch]

The thing is the real Kaveri star is the A8 7600, not the A10 models. The A8 7800 is only $119 yet thrashes any Intel APU in gaming.

Re:Of course, that would miss the point

[hairyfeet]

And anyone who went that route would frankly be short sighted and foolish and would regret it later. The simple fact is that more and more games are taking advantage of multicores and that dual core will be a bottleneck pretty quickly on any of the new games coming out. Already many of the bigger games like Bioshock infinite, Far Cry 3, Dirt 3 and the like will run better on a quad than a dual and that number is going nowhere but up.

As someone who has been selling the APUs since liano the advantages are threefold: One you can buy a nice budget system that will game OOTB and can add a dedicated GPU down the line, thus making the system more affordable to more folks, Two if you stick with AMD and buy one of the dedicated chips that support zerotech like a 7xxx chip or above you can significantly lower your power usage when you aren't gaming as zerotech will let the dedicated GPU drop down to a couple watts when not required and the APU can take care of hardware acceleration normally handled by the dedicated card, and finally Three these APUs also make great HTPC chips, with hardware acceleration for pretty much every major format out there you can get smooth 1080P playback even when using a tuner for recording or while multitasking.

As for TFA while I can't see a point in embedded 8GB of GDDR5 in a board for one of these I CAN see putting a smaller amount, say 512Mb, to give the APU a nice high bandwidth buffer to improve performance.

Re:Of course, that would miss the point

[Lonewolf666]

The 45W Kaveris are interesting, as they show a nice improvement in performance/watt - the new "sweet spot" is not in the top models but in the somewhat slower A8-7600 (3.1-3.3 GHz CPU speed).

I wonder how a 4 module (8 core) FX on that basis would perform and at which TDP. For software that scales well with many cores, it might be a good buy.

Re:AMD could do a 24 core desktop chip right now

[Anonymous Coward]

No, they don't do it because it considerably raises the cost of the chip and it doesn't help improve the "average" user's workload. Many core processors have a bunch of inherent complexity dealing with sharing information between the cores or sharing access to a common bus where the information can be transferred between processes. There are tradeoffs to improving either scenario.

Making the interconnection between cores stronger means that you have a high transistor count and many layers in silicon. Even if you do this, the scenarios in which the cores need to intercommunicate means that either the processes are being dispatched to the processor and the processor has a scheduling algorithm built in (which has its own issues), or you have a new set of instructions allowing software to dispatch code from cpu to cpu. Then race conditions and all sorts of other nonsense are inherently in the CPU itself and you have a whole bunch of problems trying to write code for the mess.

Even if you go the above route, you still have to get the data you process out of the CPU and into RAM. Then you have a bus contention problem: how can multiple cores access different sections of RAM simultaneously? Is the CPU responsible for settling a conflict when two different cores want to write to the same section of RAM at the same time? These issues are largely easily settled with only 2, 3, or 4 cores now (and currently the responsibility for preventing these scenarios is on the OS not the CPU), but they would explode with a 24 core (or more).

It's currently possible to build a CPU with several hundred cores inside (and both Intel and AMD have done it). No one has settled the issues above to make it practical, or invented the new software paradigm to make it easy to fix them.

Re:AMD could do a 24 core desktop chip right now

[symbolset]

I like to use Blender to compose my videos.

A Better Explaination At Anandtech

[Anonymous Coward]

Anandtech's writeup (which Hothardware seems to be ripping off) has a much better explanation of what's going on and why it matters.

Let me be very clear here: there's no chance that the recently launched Kaveri will be capable of GDDR5 or 4 x 64-bit memory operation (Socket-FM2+ pin-out alone would be an obvious limitation), but it's very possible that there were plans for one (or both) of those things in an AMD APU. Memory bandwidth can be a huge limit to scaling processor graphics performance, especially since the GPU has to share its limited bandwidth to main memory with a handful of CPU cores. Intel's workaround with Haswell was to pair it with 128MB of on-package eDRAM. AMD has typically shied away from more exotic solutions, leaving the launched Kaveri looking pretty normal on the memory bandwidth front.

It's also worth noting that the Anandtech article implies that AMD is still on the fence on Kaveri APUs with more memory bandwidth, and that it may be something they do if there's enough interest/feedback about it.

Re:news for nerds

[davester666]

Motherboard manufacturers see the profit margin's Apple has with RAM that can't be increased in a number of their high-end models and they want in on that action.

Re:AMD could do a 24 core desktop chip right now

[TheLink]

They don't care because a desktop with a 24 core AMD CPU is likely to be slower than a 4 core Intel CPU for most popular _desktop_ tasks which are mostly single threaded. They're already having problems competing with 8 core CPUs, adding more cores would either make their chips too expensive (too big) or too slow (dumber small cores).

Sad truth is for those who don't need the speed a cheap AMD is enough - they don't need the expensive ones. Those who want the speed pay more for Intel's faster stuff. The 8350 is about AMD's fastest desktop CPU for people who'd rather not use 220W TDP CPUs, and it already struggles to be ahead of Intel's mid range for desktop tasks: http://www.anandtech.com/bench/product/697?vs=702

A few of us might regularly compress/encode video or use 7zip to compress lots of stuff. But video compression can often be accelerated by GPUs (and Intel has QuickSync but quality might be an issue depending on implementation). The rest of the desktop stuff that people care about spending $$$ to make faster would be faster on an Intel CPU.

A server with 24 cores will be a better investment than a desktop with 24 cores.

Re:news for nerds

[SuricouRaven]

Because it's electrically so delicate that you can't keep bit sync when shoving such high frequencies through a slot connector. The price of higher bandwidth, in both the analog and digital senses.

Re:AMD could do a 24 core desktop chip right now

[symbolset]

Fine. You do it your way. I want this technological achievement so I can do it mine.

Re:AMD could do a 24 core desktop chip right now

[symbolset]

Your problem - and Intel and AMD have this problem also - is thinking the specifics of the device have something to do with its utility. It doesn't. What matters is how the thing enables people to do what they want to do, and we passed that level of utility in devices a decade ago. I can be telepresent with my children without resorting to either Intel or AMD technologies - all day. Compared to exclusively special uses of spreadsheets and slideshow programs, compatibility differentials of Office suites, that is a win.

It's about the people, not the thing. Wrap your head around that. Make it your religion.

Re:news for nerds

[hairyfeet]

Sure can. You see currently all GDDR5 is sold to GPU OEMs and they haven't used sockets on graphics card since the mid 90s.

As for TFA they sold some HD43xx boards with a small amount of dedicated RAM and since AMD makes their own board chipsets I see no reason why they couldn't sell a board or two with dedicated GDDR5 of 512Mb to say 2Gb. Lets face it guys you can only go so far with an APU before it would make more sense to have a dedicated GPU instead but with AMD having the same chips that are in the new consoles this might be a good option for a budget gaming rig.

The key will be if they can get the GDDR5 at a price that will still let the board make sense versus just buying say an HD7750 card, but if they could get the boards made so a kit would cost around $250-$270 (current their quad APU kits sell for around $200 sans HDD) this could be a really compelling option for those on a budget that want to game.

Re:AMD could do a 24 core desktop chip right now

[hairyfeet]

But the "dirty little secret" that doesn't get brought up enough frankly is a lot of those "single threaded loads" is as rigged as quack.exe was back in the day thanks to every Intel compiler made since 2002 being made to put out crippled code for any chip that Intel doesn't want to push. Oh and for those that use the "Intel just knows their own chips and optimizes for them" excuse that lie has been disproved and the proof was the last gen Pentium 3. You see the last gen P3 was curbstomping the Netburst P4s in early benchmarks, yet when the cripple compiler comes out? Suddenly the very same Netburst chips are winning by 30%!

And the bitch is that any of these so called review sites could test for rigging trivially but they won't for fear of losing Intel advertising revenue. To see if a program is rigged all one has to do is run the code on a Via CPU, Via CPUs allow one to softmod the CPUID so if you change the CPUID from "Centaur Hauls" to "Genuine Intel" and suddenly the chip scores 20%-30%+ higher on the test? Then the program has been rigged by ICC, simple as that.

All of these sites like Anandtech and Tom's have more than enough money to pick up a Via chip and keep it around for testing but they won't bite the hand that feeds so everyone should consider their tests to be tainted and as worthless as Quake tests were with the rigged drivers. It would be nice if someone would run some real tests so we could see real numbers, wouldn't be hard to do as GCC is free and there are plenty of FOSS programs like Firefox one could compile with GCC to give accurate tests but so far no review site will do this for fear of pissing off Intel. How they didn't get busted for antitrust is beyond me, this is every bit as bad as "Windows isn't done until lotus won't run" but they were allowed to bribe AMD to the tune of 2 billion to drop the lawsuit and with it the investigation.

Re:AMD could do a 24 core desktop chip right now

[K. S. Kyosuke]

The rest of the desktop stuff that people care about spending $$$ to make faster would be faster on an Intel CPU.

You mean something like this? (Or similar solutions for other languages?)

Re:news for nerds

[unixisc]

Great! Can you explain to me why "GDDR5 isn't bonded in sticks for easy motherboard socketing" ?

Reason is that they are used exclusively for the video cards/GPUs, and are not meant to be accessed directly by the CPU. In cases of integrated video on motherboards, they're not used in the first place. In case of video cards, they are soldered right on the video cards - video cards don't have slots b'cos then, you'll have video cards going into PCIe slots, and then the cards would have slots of their own, and then the height of the GDDR5 modules would potentially eliminate other motherboard slots that may be needed. In other words, configuration complications.

Short answer: GDDR5 is not meant to be socketed on to motherboards

Packaging != modular design

[unixisc]

CPUs - the last that I saw - were PGAs - except for embedded systems, I'm not sure of many BGA based CPUs. DDR3 onwards changed its packaging from TSOP to BGA due to excess pin count, where on TSOP, only a larger package would work, but then, the length of the wire bonds would become a factor in the speed of the sub-system (CPU to DDR3). Also, while TSOP is cheaper than BGA for lower pin counts, when the pin counts become comparable - ~50, the equation flips - BGAs become cheaper than 56 pin TSOPs. As a result, memory started getting packaged in BGA, which were already in use by flash memory, especially in portable applications such as cell phones, PDAs and so on.

But even those DDR3 modules are available on sticks, even in their BGA packages.